End-to-End Light Field Image Compression with Sparse Reconstruction at Low Bit Rates

Unlike traditional light field image (LFI) compression techniques, end-to-end LFI compression achieves better compression efficiency by enabling joint optimization of various modules. However, due to the unique structure of LFI, LFI transmission at low bit rates presents significant challenges. Therefore, an end-to-end LFI compression network combined with sparse reconstruction for low bit rates is proposed in this paper. Firstly, to reduce the data volume for transmitting LFI under limited bit rate conditions, an end-to-end sparse reconstruction LFI compression framework is constructed, in which only a subset of sub-aperture images (SAIs) are compressed, while the remaining SAIs are synthesized through a reconstruction network. Secondly, a hybrid spatial-angular attention module (HSAAM) is designed specifically for LFI. It considers the spatial and angular correlations of LFI from SAIs, micro-lens image, and epipolar plane images, and obtains compact LFI features through a large kernel attention mechanism for compression. The experimental results show that compared to pseudo-video-based and multi-view video-based compression methods, the proposed method achieves better compression efficiency at low bit rates, along with higher subjective visual quality and structural consistency.